{
  "schema": 1,
  "id": "change-detection-v1",
  "status": "Prospective local protocol; sources and protocol hashed before fresh admission. Development uses previously examined action tables. Not public preregistration.",
  "question": "Can feedback-only change detection and selective forgetting improve individual adaptation without damaging stable performance?",
  "seeds": [
    31001,
    31042,
    31083,
    31124,
    31165,
    31206,
    31247,
    31288,
    31329,
    31370,
    31411,
    31452,
    31493,
    31534,
    31575,
    31616
  ],
  "development_seeds": [
    22001,
    22038,
    22075,
    22112
  ],
  "groups": [
    "solo-vanguard",
    "solo-flanker",
    "solo-ranger"
  ],
  "candidates_per_seed": 32,
  "training_rounds": 32,
  "test_rounds": 64,
  "detector": {
    "window": 4,
    "threshold": 0.2,
    "confirmations": 2
  },
  "minimum_qualified_seeds": 12,
  "minimum_gain": 1,
  "stable_harm_margin": 0.5,
  "maximum_false_alarms_per_100": 1,
  "primary_uncertainty": {
    "resamples": 24000,
    "seed": 260031,
    "interval": 0.9958333333333333
  },
  "secondary_uncertainty": {
    "resamples": 5000,
    "seed": 260032,
    "interval": 0.95
  },
  "admission": "Unchanged opportunity-v2 admission: enumerate all 32 candidates and both legal tactics in A and B; disjoint optima, >=12 HP and >=0.15 reward separation both directions, all feedback observed. Select first qualified index per role/seed before learning. Keep every rejection; no replacement seed.",
  "schedules": "Shared 32-encounter warmup from stationary 90% A / 10% B. All arms fork identical memory, counters and RNG. Each test history is 64 encounters: abrupt changes to 90% B; gradual ramps B probability from 10% to 90% over first 32, then stays; recurring is 90% B for 32 then 90% A for 32; stable remains 90% A for all 64. Bernoulli draws keyed by seed, role, history and round and shared by arms. Drift means changing encounter mixture, not continuously morphing movement within a fight.",
  "arms": "Retained online; exact warmup-frozen; detector resetting both action cells in detected role/context; detector resetting only detected action cell; signaled reset of all solo cells before an offline expected-damage optimum changes. Ties keep the previous optimal action. Signaled timing uses the true mixture and full action tables and is a privileged reference, not an achievable ceiling.",
  "detector_rule": "After commit, compare four older with four newer observed rewards for that action. Two consecutive same-direction differences with magnitude >=0.2 trigger. Observation gaps do not fill buffers. Keep newer four after alarm, reset streak; maximum eight rewards and two scalars per configured action. Warmup fills buffers but does not arm interventions. Reset scores and sample counts to zero only; preserve all other snapshot fields. Changes affect only subsequent decisions. No strategy labels, schedule, future reward, or oracle enter this detector.",
  "primary": "Six role-by-detector comparisons: mean damage difference against retained online, equally averaging all 64 rounds in each of the three changing histories within each seed. Advancement requires >=12 qualified seeds, >=1 HP mean gain, adjusted lower CI >0, adjusted stable-history lower CI >-0.5 HP, and <=1 observed false alarm per 100 stable fights. Twelve bootstrap intervals (six gains and six noninferiority checks) use nominal 99.5833% coverage, Bonferroni 0.05/12. This is an approximate small-sample engineering screen. Stable alarm cap is a descriptive gate, not a population rate confidence bound. All other comparisons exploratory. No production promotion.",
  "detection": "An event starts at test round 1 for abrupt/gradual, and at 1 and 33 for recurrence. First autonomous alarm in each event window is an operational detection; count all missed windows and all later alarms. Delay counts completed encounters, earliest 1. Misses right-censored at window length and encoded length+1 only for restricted-check summaries. An alarm in a window does not prove it was caused by the distribution change. Every stable-history alarm is false relative to the stationary distribution, including B outliers. Record all alarms, not just first.",
  "recovery": {
    "threshold": 0.35,
    "consecutive": 3,
    "definition": "At each offline expected-damage-optimum crossover, measure first ending round of three consecutive pre-choice distributions with expected regret <=35% of current distribution best-worst gap. Expected regret integrates both true A/B mixture and actual action probabilities. Ties have zero regret. Report initially displaced separately, right-censor at next crossover/end; encode horizon+1 for restricted checks, never discard misses. Secondary measure; criterion can already hold before learning."
  },
  "regret": "Realized regret uses best legal action in the realized A/B state minus actual damage. Distribution expected regret instead integrates both state mixture and action distribution; it compares with the best action before the latent state is known. These oracles are different and used offline only. Mean damage differences against identical schedules equal differences in realized regret.",
  "scope": "Fresh parameter/world seeds for familiar authored controllers; deliberately action-sensitive pairs. Stochastic ordering of two deterministic conditional worlds creates stationary outcome noise. This is not human behavior, arbitrary maps, new behavior families, team coordination, or continuous within-fight adaptation. Eight reward slots per action is a heuristic engineering detector, not a calibrated statistical change-point test. Formal trials execute actual archived Rust/Wasm choices and archived combat physics; development may reuse validated outcome tables and is labeled separately.",
  "demonstration": {
    "seed": 31001,
    "selection": "First reserved seed, first qualifying index per role; show absence if none. All arms and histories available. Initial displayed test round 8, irrespective of result."
  },
  "development_policy": "One declared detector rule, no parameter sweep. Development on first qualified previously examined cases in four listed seeds. Preserve all cached-policy outcomes and sources; do not count cached rewards as new physical encounters. Fix correctness defects before freezing. Reserved outcomes do not authorize retuning this version."
}